Fix Issue 11025 - case-insensitive backreferences in std.regex

[niy-ati]

Summary

• Under /i, IR.Backref compared captures with ==, so (.)\1 matched BB but not Aa.
• Compare with simple case folding when RegexOption.casefold is set (backtracking, Thompson, and ctRegex).
• Add regression vectors for issue 11025.

Fixes #11025

Test plan

• Added TestVectors for Aa / aA / BB / non-match / multi-char / (?i).
• CI

[DmitryOlshansky]

What I meant was not using sicmp as is but getting its internal compare function and using that to compare two dchar case-insensitively. Plain call to sicmp brings its own overheads which are not acceptable.

[niy-ati]

Thanks for the clarification ,
I’ll add/reuse the internal simple-case per-dchar compare from sicmp (trie tables) and use that from equalCasefold in std.regex.internal.ir, without calling sicmp itself. I’ll refactor so we don’t duplicate the trie logic if that fits Phobos style.

[DmitryOlshansky]

LGTM

[kubo39]

Literals are folded at parse time using the flag active at their position, but a backref here uses one global flag for the entire pattern.
That diverges as soon as (?i)/(?-i) toggles case-insensitivity around the backref:

// over-match: (?i) appears *after* the backref, so \1 should stay case-sensitive
assert( matchFirst("Aax", regex(`(.)\1(?i)X`)).empty);   // fails: matches "Aax"

// under-match: casefold is on at \1 but cleared before X
assert(!matchFirst("AaX", regex(`(?i)(.)\1(?-i)X`)).empty); // fails: no match

[DmitryOlshansky]

The way out is adding a new opcode for caseinsensitive backreference

[niy-ati]

Addressed the inline (?i) / (?-i) issue by introducing IR.BackrefI: the parser records whether casefold is active at each backref site and emits Backref vs BackrefI accordingly. Matchers no longer use re.flags for backref comparison. Added tests for (.)\1(?i)X / (?i)(.)\1(?-i)X alongside the existing #11025 vectors. Case-insensitive comparison still uses simpleCaseCmp via equalCasefold on the BackrefI path only.

@kubo39 , backrefs now follow the same per-position casefold as literals (Backref / BackrefI at parse time). @DmitryOlshansky , implemented the separate case-insensitive backref opcode as suggested.

[DmitryOlshansky]

LGTM

[niy-ati]

@rikkimax whenever you have a moment pls review it, open to address anything else.

LGTM

#11070 (comment)

[rikkimax]

@rikkimax whenever you have a moment pls review it, open to address anything else.

I have nothing to add std.regex isn't really my area. But I do get to say that I'd love for someone to finish case handling in std.uni with the same vigor!

Would be very helpful to get that finished before PhobosV3 so I don't have to deal with that.

[niy-ati]

FreeBSD 14.4 failed during pkg install (mpdecimal fetch / disk write) before the suite ran. Other checks are green , could someone re-run the failed FreeBSD job?

[0xEAB]

@niy-ati see also dlang/ci#497

[0xEAB]

• LGTM
• Passes testsuite.
  • Does not remove or alter any tests.
  • Adds new tests.
• Has code owner’s approval: #11070 (comment)
• I’ve added nitpick format changes.